CLARA: Composable Learned Assured Reasoning Architecture Demonstration Suite for DARPA TA1 Assured Reasoning

The Problem

AI systems deployed in safety-critical domains -- military targeting, autonomous vehicles, medical diagnostics, critical infrastructure -- need more than accuracy. They need formal assurance guarantees that certain behaviors will never occur, regardless of what the learned model predicts. A 99.9% accurate targeting system that occasionally authorizes strikes on protected sites is not an acceptable system. It is a liability.

Traditional machine learning provides statistical accuracy but offers no hard safety guarantees. A deep neural network can classify targets with high precision, but nothing in its architecture prevents it from producing catastrophically wrong outputs on adversarial or out-of-distribution inputs. Conversely, traditional rule-based systems provide formal safety properties -- if the rules are correct, the system is safe -- but they cannot learn from data, generalize to novel situations, or improve with experience.

The gap is architectural: how do you compose a learned ML model with a formal safety rule system such that the safety guarantees are provably preserved regardless of ML quality? This is not a matter of adding a filter after the fact. The composition must be structural, auditable, and domain-independent.

The CLARA Approach

CLARA solves this through constraint-objective separation. Safety constraints (Levels 1-4: Law of Armed Conflict, No-Strike Lists, Rules of Engagement, Collateral Damage Estimation) are encoded as immutable, write-protected rules in a directed acyclic graph (DAG). Mission objectives (Levels 5-6: tactical priorities, user preferences) are encoded as separate, mutable rules that can adapt to operational context. The architecture guarantees that no objective can override a constraint -- this is enforced structurally, not by convention.

The rule DAG uses namespace isolation to prevent interference between rule categories. Each namespace (e.g., loac, nsl, roe, cde) is independently verifiable, and the composition engine proves non-interference at every decision point. SHA-256 hash verification confirms that constraint rules remain invariant across all mission profiles and operational modes.

Critically, CLARA is ML-agnostic. Any classifier -- a convolutional neural network, logistic regression, or an ensemble -- composes through the same DAG with the same safety guarantees. A bad classifier does not produce unsafe decisions; it produces conservative decisions (more targets held for human review). The same rules engine that enforces constraints at inference time also produces the cost matrix used during ML training, closing the loop between assurance and learning.

Demonstration Suite

Key Results